Xiaomi launches three new models: 1 trillion-parameter flagship and 10x speed variant
The Chinese technology giant has unveiled the MiMo-V2.6 series with three models targeting agentic coding and long-horizon tasks directly. The flagship, MiMo-V2.6-Pro, boasts more than 1 trillion parameters, placing it in the same weight class as the flagship models from OpenAI, Anthropic and Google. Alongside it sits MiMo-V2.6-Flash, a 309 billion-parameter mixture-of-experts architecture with only 15 billion active parameters per forward pass. The headline act, however, is MiMo-V2.6-Pro-UltraSpeed, which promises the same quality as the Pro model at roughly ten times the token emission rate.
The flagship: 1 trillion parameters for a 1 million-token context window
MiMo-V2.6-Pro is built as a 1T+ parameter foundation model with a 1 million-token context window, enabling it to ingest entire codebases, long technical documents and extended conversation histories without resorting to RAG or chunking. The model supports native multimodal capabilities — meaning it processes text, images and presumably video in a single representation space, without external modules stitching together separate outputs. Xiaomi defines the primary purpose as "agentic workflows": tasks where the model plans, executes, verifies and self-corrects autonomously over a long horizon, rather than simply answering single queries.
Flash: MoE efficiency with only 15B active
MiMo-V2.6-Flash offers a leaner alternative: 309 billion total parameters in an MoE architecture where just 15 billion activate per forward pass. The practical implication is significantly lower inference cost relative to the flagship while retaining a substantial portion of its capabilities — a pattern familiar from models such as Mixtral or DeepSeek-MoE, which enable deployment on more modest hardware or higher throughput for the same compute budget. Xiaomi has not published comparative benchmarks between Flash and Pro, so the extent of any quality trade-off remains an open question.
UltraSpeed: same checkpoint, entirely different throughput
The most technically intriguing announcement is MiMo-V2.6-Pro-UltraSpeed. Xiaomi stresses that this is the exact same checkpoint as the Pro model — identical weights, identical quality — that has undergone aggressive inference optimisation yielding roughly ten times the output token throughput. If the claim holds under real-world testing, it represents a significant engineering achievement: speed-ups of this magnitude typically require aggressive quantisation that degrades quality, or techniques such as speculative decoding that add complexity and behavioural edge cases. The fact that the 1 million-token window and multimodal capabilities are preserved intact suggests deep kernel-level and memory optimisation work, not merely sampling tricks.
No benchmark numbers, no peer review
What is missing from the announcement stands out as much as what is present. Xiaomi has not released results on standard benchmarks such as SWE-bench, HumanEval, MMLU, GPQA or MMMU, nor has it shown direct comparisons against GPT-4o, Claude 3.5 Sonnet or Gemini 1.5 Pro. No technical report or pre-print details the training methodology, data mixture, or the evaluation protocol behind claims of "top-tier performance in coding, visual, general and research scenarios." Absent that transparency, the technical community will treat the claims as vendor assertions, not verified measurements. Experience shows that gaps between launch presentations and production performance can be wide, especially in long-horizon agentic tasks where errors compound.
Context: Xiaomi building its own AI stack
The launch fits a broader push by Xiaomi to reduce dependence on closed models from US rivals and build a vertical channel from hardware — smartphones, vehicles, IoT — through operating system to proprietary foundation model. A model with a 1 million-token window and strong agentic capabilities could power on-device coding assistants across Xiaomi's product line, automate QA testing, and analyse long logs without sending sensitive data to external clouds. For developers working with Xiaomi's supply chain or building applications for its ecosystem, availability of such a model — if released as open weights or via a competitive API — could become a relevant alternative to Western models, particularly in scenarios demanding data sovereignty or low-cost inference at scale.